COMMENTARies | special feature
Is sequencing enlightenment
ending the dark age of the
transcriptome?
© 2009 Nature America, Inc. All rights reserved.
Piero Carninci
Sequencing-based technologies for RNA discovery are playing a
key role in deciphering the transcriptome and hold the potential to
provide us with a census of RNAs and their functions.
A decade ago, before the appearance of the
first draft of the human and other genomes,
genes were identified via expressed sequencing tags. In the pregenome era, the main
goal was to identify mRNAs produced by
protein-coding genes. The computational
tools used to predict these genes were far
from accurate and required experimental
data to support gene predictions. Contrary
to expectations, just at the turn of the new
millennium, initial genome studies suggested that the total number of genes was much
lower than estimated earlier by analyses of
expressed sequencing tag data1, giving an
upper bound of approximately 25,000 protein-coding mammalian genes.
The first doubts about this estimate
came from the analysis of full-length
cDNAs, which showed that in addition to
protein-coding genes, there is an even larger number of RNA transcripts that do not
encode proteins and are instead defined as
noncoding RNAs (ncRNAs)2. Additionally,
the analysis of 1% of the human genome
by the ENCODE (Encyclopedia of DNA
Elements) Consortium not only suggested
that 93% of the genome is transcribed but
also revealed an unprecedented number of
splice variants. These data put the number of predicted proteome variants at least
fourfold above the number of genes3.
Indeed, an average cell contains at least
300,000 mRNA molecules and perhaps
even more if we consider rare ncRNAs or
RNAs restricted to specific cell compartments. Different cells express different sets
of mRNAs and ncRNAs, and their expression level spans 4–5 orders of magnitude
from the highly to the rarely expressed
RNA. Classical Sanger (dideoxy) sequencing of full-length cDNA, even if implemented with strategies to subtract the
known genes to more efficiently search
for new RNAs4, still requires the cloning
of individual cDNAs before analysis of the
clones; the largest full-length sequencing
efforts so far have been limited to about
100,000 cDNAs2. The main limitation of
this approach is the need for very extensive handling and sequencing of individual
cDNA clones. The advantage is that once
the full-length cDNAs are sequenced, they
not only serve to identify mRNAs, ncRNAs
and their isoforms, but they can also be
used in functional assays.
Tiling DNA microarrays, which probe
the whole nonrepetitive part of the genome
at high resolution5, are faster alternatives
for transcript detection in that they instantaneously allow the comprehensive assessment of mRNA and ncRNA expression.
Together with tagging technologies6 that
identify novel RNA start and termination
sites, tiling arrays have been instrumental in revealing a seemingly overwhelming degree of transcriptome complexity.
However, they suffer from a lack of sensitivity in detecting rare transcripts and
Piero Carninci is at the Omics Science Center, RIKEN Yokohama Institute, Yokohama, Kanagawa, Japan.
e-mail: [email protected]
from background owing to cross-hybridization of highly related sequences, making the analysis of repeat regions impossible. Additionally, tiling arrays cannot
be used to identify new splice junctions:
they will identify exons but do not provide
their connections. Even arrays dedicated
to the detection of predicted or known
splice variants cannot detect novel splice
sites that have not yet been experimentally
identified or predicted in silico.
The arrival of second-generation
sequencing has advanced unbiased transcriptome studies. The technology, known
as RNA-sequencing (or RNA-seq) is procedurally quite simple: RNA is converted to
a cDNA library without a lengthy cloning
procedure, and a single cDNA preparation
can yield over hundreds of millions of
short reads that are then computationally
aligned to the genome. Sequences mapping to a single exon are generally unambiguously assigned to the corresponding
gene, often at a coverage of hundreds of
sequences for each RNA. Chances are good
that even rare RNAs can be identified in
the library, although full assembly from the
short reads is still unlikely for transcripts
present in one copy per cell at the current
coverage. With increased sequencing depth
this may change, and it may become possible that RNA-seq will sensitively detect
the whole transcriptome in a biological
sample.
Indeed, a growing number of studies
published in this journal and elsewhere7–11
have used RNA-seq to comprehensively
investigate transcription. One of the obvious technical advantages of RNA-seq is
the avoidance of spurious, false positive
detection of RNA, an inherent problem of
microarrays owing to cross-hybridization.
RNAs that are produced by highly similar
members of paralogous genes present an
alignment challenge, but recently developed strategies also allow for the mapping
of reads to splice sites or to paralogous
sequences9. With read length increasing,
it will become even easier to identify the
exact location of most sequence reads.
In one of the first uses of RNA-seq,
Mortazavi and colleagues developed an
nature methods | VOL.6 NO.10 | OCTOBER 2009 | 711
special feature | COMMENTARies
Genome
RNA structure
1
© 2009 Nature America, Inc. All rights reserved.
RNA-seq
2
3
4
5
?
Start site
?
Figure 1 | Challenges in assembling RNA-seq data. A hypothetical locus often produces multiple RNAs
(top), which may affect assembly of RNA-seq into an unambiguous gene model. A given annotated
protein-coding gene (the known annotated region is represented by back boxes, and known exons
are numbered 1–5) produces protein-coding mRNAs (gray boxes and arrrows) and other noncoding
RNAs (large red arrows). RNA-seq produces multiple overlapping signals (bottom), represented by
black and red tiled arrows. Although RNA-seq allows reconstruction of exon structure and neighboring
exon-exon junction definition, the connectivity of exons as well as all the splice combinations of
individual transcripts may not be fully reconstructed. By looking at RNA-seq data, the reconstruction of
connectivity in the various transcripts between distant exons is challenging. Additionally, RNAs have
multiple transcription start sites (middle, small red arrows), identified, for instance, by cap analysis
gene expression (CAGE tags), which cannot be detected efficiently by RNA-seq.
RNA-seq method in which oligo-dT–selected mouse mRNAs are fragmented before
synthesis of random-primed, first-strand
cDNA9. This RNA fragmentation proved
essential for uniform coverage of all exons.
Random priming performed on full-length
mRNA would show a strong preference to a
few particular positions on the mRNAs and
thus bias coverage. Although the method
was limited to the short 25-nucleotide reads
available in the first version of the Illumina
GA sequencer, the authors impressively
provided a clear pattern of novel splice sites.
These findings explore the world of protein-coding mRNAs, as noncoding RNAs
often lack poly(A) tails, and their data suggest that a good fraction of protein-coding
exons has already been identified because
only a few percent of the tags mapped to
intergenic regions.
In parallel, Sean Grimmond and colleagues went for even larger transcriptome coverage in human embryonic stem
cells and embryoid bodies using Applied
Biosystems’ sequencing by ligation
(SOLiD) technology10. They reasoned that
RNAs are often transcribed bidirectionally
and often overlap to form sense-antisense
pairs. To capture these sense-antisense
RNA pairs, they devised a template switch
strategy that maintains the 5′-3′ directionality of the original RNAs in the shotgun
library: the reverse transcriptase adds
(albeit at low efficiency) some extra bases
at the end of the first-strand cDNA, so the
second-strand cDNA can be primed from
the overhang. Another important procedural step was that they subtracted ribosomal RNAs from the total RNA preparation before cDNA library preparation;
otherwise ribosomal RNAs would comprise a large part of the library, affecting
sequencing yield. The authors showed that
sense-antisense pairs are more concentrated at the 3′ end of genes.
A remaining challenge is the detection of
rare RNAs expressed only in a few subtypes
of cells. This is a key issue in analyzing complex tissues, such as in the brain and those
formed during embryonic development.
A first step toward meeting this challenge
has been made by Surani and colleagues,
who prepared seven libraries from the
limited amount of RNA in single mouse
oocytescells in an early developmental
stage11. Owing to the protocol, which is
based on oligo-dT priming followed by
A-tailing of cDNA and subsequent cDNA
712 | VOL.6 NO.10 | OCTOBER 2009 | nature methods
fragmentation before sequencing, the
direction of the RNA is lost. There are no
protocols yet for the removal of ribosomal
RNA when only a few nanograms of RNA
are available. Also, oocytes, though technically single cells, are much larger than
an average cell, which leaves room for the
development of technologies to miniaturize single-cell sequencing–based assays.
It is certainly reassuring that different
sequencing and library preparation technologies have pointed unambiguously
to similar levels of complexity and all
have provided solutions for tackling this
complexity.
Indeed, RNA-seq is contributing to the
comprehensive analysis of splice events and
novel splicing combinations. Even with relatively ‘shallow’ coverage, such as less than
10 million 27-nucleotide short sequences7,
researchers found more than 2,000 new
exons for each cell line derived from kidney
and B cells. Deeper sequencing of longer
fragments allowed investigators to detect
millions of splice sites and ~90,000–145,000
novel splice site candidates per tissue8–11.
This clearly supports the notion that multiexon genes are usually alternatively spliced,
but these studies have also demonstrated
that alternative splicing is restricted to only
a fraction of exons.
RNA-seq also deals much better with
RNA derived from repeat elements 10 ,
which have been previously excluded from
arrays because of cross-hybridization.
Notably, even 20-nucleotide sequence tags
have been suitable for detection of widespread expression of retrotransposons12.
Have we addressed all the challenges yet?
A final unambiguous assembly of different
splice isoforms is still outstanding because
the connectivity between distant exons
that do not share sequencing reads cannot
be unequivocally predicted (Fig. 1). The
ENCODE analysis working group is currently developing computational methods
to assemble RNA-seq reads according to
RNA models with accurate splice sites and
to predict the most likely splice variants
based on the expression of each exon. At the
same time, sequencing technology is constantly improving, reads are longer—and
thus the assembly of full mRNA sequences
is becoming simpler—and greater support
is being provided with the second- and
third-generation sequencing technologies.
The Roche 454 Life Sciences sequencer can
produce over a million sequences around
500 nucleotides long; this will be helpful
© 2009 Nature America, Inc. All rights reserved.
COMMENTARies | special feature
for assembling splicing sites. The ABI
SOLiD and the Illumina GAIIX have not
only increased the sequencing length to
50 and 75 bases, respectively, but have also
developed methods for sequencing from
both ends of the cDNA fragments to help
in connecting more distant exons.
Other challenges of RNA-seq are how to
distinguish the various start and end sites of
RNAs. It is becoming evident that there are
often multiple overlapping RNAs encoded
from the same genome region, and intronderived RNAs are recycled to produce
functional ncRNAs such as microRNAs.
Another source of complexity comes from
the secondary processing of mRNAs, which
produces shorter, likely functional, RNAs.
Thus, protein-coding genes are associated
with a plethora of short ncRNAs, including
short RNAs associated with promoters13,
transcripts arising around termination sites
and even exons. A fraction of these RNAs
are produced by a novel cleavage and recapping mechanisms, resulting in capped RNAs
that start in the middle of coding exons or
in untranslated regions. These naturally
truncated RNAs are likely to be ncRNAs
that overlap larger mRNAs13. Another complication arises from the broad nature of
many promoters14, which produce various
capped RNAs from multiple transcription
start sites. Technologies that identify the
cap structure in such mixtures are needed
to distinguish the RNA fragments obtained
by RNA-seq. At present RNA-seq does not
perform well at unambiguously identifying
transcription start sites, and RNA-seq protocols need improvement to simultaneously
decipher the long, short and capped RNAs
so the RNAs’ function can be assessed.
Some of the third-generation sequencers
such as those from Pacific Biosciences and
Oxford Nanoporewhich will be able to
read thousands of nucleotides15 of single
cDNAsmay ultimately meet these challenges: their long sequences will quantitatively represent complete RNAs, and the
use of tags and linkers that mark cap sites
and other modifications will allow an allin-one determination of transcriptome
structure, including start and termination sites and the mapping of regulatory
elements such as promoters. The accurate
sequence of coding sequences will also help
directed cloning of open reading frames
in experimental verification of alternative
splice isoforms16,17.
Although many challenges are ahead,
the direction is becoming clearer, and I am
beginning to wonder if the dark age of the
transcriptome is giving way to rays of light.
1. Liang, F. et al. Nat. Genet. 25, 239–240 (2000).
2. Carninci, P. et al. Science 309, 1559–1563
(2005).
3. ENCODE Project Consortium et al. Nature 447,
799–816 (2007).
4. Carninci, P. et al. Genome Res. 13, 1273–1289
(2003).
5. Kapranov, P. et al. Science 316, 1484–1488
(2007).
6. Harbers, M. & Carninci, P. Nat. Methods 2,
495–502 (2005).
7. Sultan, M. et al. Science 321, 956–960 (2008).
8. Pan, Q., Shai, O., Lee, L.J., Frey, B.J. &
Blencowe, B.J. Nat. Genet. 40, 1413–1415
(2008).
9. Mortazavi, A., Williams, B.A., McCue, K.,
Schaeffer, L. & Wold, B. Nat. Methods 5, 621–
628 (2008).
10. Cloonan, N. et al. Nat. Methods 5, 613–619
(2008).
11. Tang, F. et al. Nat. Methods 6, 377–382 (2009).
12. Faulkner, G.J. et al. Nat. Genet. 41, 563–571
(2009).
13. Fejes-Toth, K. et al. Nature 457, 1028–1032
(2009).
14. Carninci, P. et al. Nat. Genet. 38, 626–635
(2006).
15. Turner, D.J., Keane, T.M., Sudbery, I. & Adams,
D.J. Mamm. Genome 20, 327–338 (2009).
16. Djebali, S. et al. Nat. Methods 5, 629–635
(2008).
17. Salehi-Ashtiani, K. et al. Nat. Methods 5,
597–600 (2008).
Engineered fluorescent proteins:
innovations and applications
Michael W Davidson & Robert E Campbell
Despite expansion of the fluorescent protein and optical
highlighter palette into the orange to far-red range of the visible
spectrum, achieving performance equivalent to that of EGFP has
continued to elude protein engineers.
Evolving proteins, evolving tools
During the past decade and a half, intrinsically fluorescent proteins have been under
intense evolutionary pressure for ‘fitness’,
not in the wild, but rather for utility in livecell imaging experiments. This unnatural
course of evolution has occurred on the
benches of protein engineers around the
world who have helped to drive progress in
the ever-expanding repertoire of fluorescence imaging technologies.
An underlying theme that has guided
advancements in fluorescent protein engineering is that, all other factors being equal,
redder is better. It is generally accepted that
excitation with longer-wavelength light
entails less phototoxicity for the cells or
tissue being examined and decreased autofluorescence and scattering. These desirable
factors mean that red-shifted fluorophores
generally provide improved contrast (owing
to decreased background fluorescence) and
superior performance in whole-organism
imaging (owing to higher tissue ‘transparency’). Early efforts to engineer red-shifted
Aequeorea victoria GFP (avGFP) variants
led to the development of enhanced GFP
(EGFP) and yellow fluorescent proteins
with emission maxima at approximately 507
nm and 529 nm, respectively (versus 508 nm
for wild type)1.
For a time, however, it appeared that
fluorescent protein engineering had hit a
‘yellow’ wall in efforts to red-shift fluorescence emission. Fortuitously, this barrier
had already been surmounted by natural
evolution, as was revealed in October 1999
with a report that the Discosoma sp. mushroom anemone harbored a fluorescent protein homolog, commonly known as DsRed,
emitting in the orange-red region (583
nm)2. Counterbalancing this favorable shift
to the red were several undesirable properties, including oligomerization, ‘contamination’ by a green component and sluggish
chromophore development, which dampened some of the initial enthusiasm.
The discovery of DsRed (and other
Anthozoa fluorescent proteins of various hues) had a twofold impact on the
Michael W. Davidson is at the National High Magnetic Field Laboratory and Department of Biological
Science, Florida State University, Tallahassee, Florida, USA. Robert E. Campbell is at the University of Alberta,
Department of Chemistry, Edmonton, Alberta, Canada.
e-mail: [email protected]
nature methods | VOL.6 NO.10 | OCTOBER 2009 | 713